freedreno/ir3: two pass register allocation
authorRob Clark <robdclark@chromium.org>
Tue, 22 Oct 2019 21:00:09 +0000 (14:00 -0700)
committerMarge Bot <eric+marge@anholt.net>
Sat, 1 Feb 2020 02:40:22 +0000 (02:40 +0000)
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3569>

src/freedreno/ir3/ir3_compiler_nir.c
src/freedreno/ir3/ir3_ra.c

index 52a049a4123b02953abc9a0a3cbd9618c81ce917..a7536cfecd2a73c7ac0fcbe8b0a887270dbd2fc2 100644 (file)
@@ -3479,8 +3479,6 @@ ir3_compile_shader_nir(struct ir3_compiler *compiler,
                goto out;
        }
 
-       ir3_debug_print(ir, "AFTER RA");
-
        ir3_postsched(ctx);
        ir3_debug_print(ir, "AFTER POSTSCHED");
 
index ff357e91fb7f626a835a980e6a06d5cfb5ea2048..73fc10e39f2cfc295ad7300c17e6b19797ea70a2 100644 (file)
  * subsequent partial writes to r0.xy.  So the 'add r0.z, ...' is the
  * defining instruction, as it is the first to partially write r0.xyz.
  *
- * Note i965 has a similar scenario, which they solve with a virtual
- * LOAD_PAYLOAD instruction which gets turned into multiple MOV's after
- * register assignment.  But for us that is horrible from a scheduling
- * standpoint.  Instead what we do is use idea of 'definer' instruction.
- * Ie. the first instruction (lowest ip) to write to the variable is the
- * one we consider from use/def perspective when building interference
- * graph.  (Other instructions which write other variable components
- * just define the variable some more.)
+ * To address the fragmentation that this can potentially cause, a
+ * two pass register allocation is used.  After the first pass the
+ * assignment of scalars is discarded, but the assignment of vecN (for
+ * N > 1) is used to pre-color in the second pass, which considers
+ * only scalars.
  *
  * Arrays of arbitrary size are handled via pre-coloring a consecutive
  * sequence of registers.  Additional scalar (single component) reg
@@ -289,7 +286,7 @@ ir3_ra_alloc_reg_set(struct ir3_compiler *compiler)
                for (unsigned i = 0; i < half_class_count; i++) {
                        /* NOTE there are fewer half class sizes, but they match the
                         * first N full class sizes.. but assert in case that ever
-                        * accidentially changes:
+                        * accidentally changes:
                         */
                        debug_assert(class_sizes[i] == half_class_sizes[i]);
                        for (unsigned j = 0; j < CLASS_REGS(i) / 2; j++) {
@@ -334,6 +331,13 @@ struct ir3_ra_ctx {
 
        struct ir3_ra_reg_set *set;
        struct ra_graph *g;
+
+       /* Are we in the scalar assignment pass?  In this pass, all larger-
+        * than-vec1 vales have already been assigned and pre-colored, so
+        * we only consider scalar values.
+        */
+       bool scalar_pass;
+
        unsigned alloc_count;
        /* one per class, plus one slot for arrays: */
        unsigned class_alloc_count[total_class_count + 1];
@@ -411,6 +415,12 @@ get_definer(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr,
        struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
        struct ir3_instruction *d = NULL;
 
+       if (ctx->scalar_pass) {
+               id->defn = instr;
+               id->off = 0;
+               id->sz = 1;     /* considering things as N scalar regs now */
+       }
+
        if (id->defn) {
                *sz = id->sz;
                *off = id->off;
@@ -428,7 +438,7 @@ get_definer(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr,
                /* note: don't use foreach_ssa_src as this gets called once
                 * while assigning regs (which clears SSA flag)
                 */
-               foreach_src_n(src, n, instr) {
+               foreach_src_n (src, n, instr) {
                        struct ir3_instruction *dd;
                        if (!src->instr)
                                continue;
@@ -590,12 +600,36 @@ ra_block_name_instructions(struct ir3_ra_ctx *ctx, struct ir3_block *block)
                if (id->defn != instr)
                        continue;
 
+               /* In scalar pass, collect/split don't get their own names,
+                * but instead inherit them from their src(s):
+                *
+                * Possibly we don't need this because of scalar_name(), but
+                * it does make the ir3_print() dumps easier to read.
+                */
+               if (ctx->scalar_pass) {
+                       if (instr->opc == OPC_META_SPLIT) {
+                               instr->name = instr->regs[1]->instr->name + instr->split.off;
+                               continue;
+                       }
+
+                       if (instr->opc == OPC_META_COLLECT) {
+                               instr->name = instr->regs[1]->instr->name;
+                               continue;
+                       }
+               }
+
                /* arrays which don't fit in one of the pre-defined class
                 * sizes are pre-colored:
                 */
                if ((id->cls >= 0) && (id->cls < total_class_count)) {
-                       instr->name = ctx->class_alloc_count[id->cls]++;
-                       ctx->alloc_count++;
+                       /* in the scalar pass, we generate a name for each
+                        * scalar component, instr->name is the name of the
+                        * first component.
+                        */
+                       unsigned n = ctx->scalar_pass ? dest_regs(instr) : 1;
+                       instr->name = ctx->class_alloc_count[id->cls];
+                       ctx->class_alloc_count[id->cls] += n;
+                       ctx->alloc_count += n;
                }
        }
 }
@@ -660,6 +694,27 @@ ra_name(struct ir3_ra_ctx *ctx, struct ir3_ra_instr_data *id)
        return __ra_name(ctx, id->cls, id->defn);
 }
 
+/* Get the scalar name of the n'th component of an instruction dst: */
+static int
+scalar_name(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr, unsigned n)
+{
+       if (ctx->scalar_pass) {
+               if (instr->opc == OPC_META_SPLIT) {
+                       debug_assert(n == 0);     /* split results in a scalar */
+                       struct ir3_instruction *src = instr->regs[1]->instr;
+                       return scalar_name(ctx, src, instr->split.off);
+               } else if (instr->opc == OPC_META_COLLECT) {
+                       debug_assert(n < (instr->regs_count + 1));
+                       struct ir3_instruction *src = instr->regs[n + 1]->instr;
+                       return scalar_name(ctx, src, 0);
+               }
+       } else {
+               debug_assert(n == 0);
+       }
+
+       return ra_name(ctx, &ctx->instrd[instr->ip]) + n;
+}
+
 static void
 ra_destroy(struct ir3_ra_ctx *ctx)
 {
@@ -674,7 +729,7 @@ __def(struct ir3_ra_ctx *ctx, struct ir3_ra_block_data *bd, unsigned name,
        /* defined on first write: */
        if (!ctx->def[name])
                ctx->def[name] = instr->ip;
-       ctx->use[name] = instr->ip;
+       ctx->use[name] = MAX2(ctx->use[name], instr->ip);
        BITSET_SET(bd->def, name);
 }
 
@@ -714,30 +769,10 @@ ra_block_compute_live_ranges(struct ir3_ra_ctx *ctx, struct ir3_block *block)
                }
        }
 
-
        foreach_instr (instr, &block->instr_list) {
                struct ir3_instruction *src;
                struct ir3_register *reg;
 
-               /* There are a couple special cases to deal with here:
-                *
-                * split: used to split values from a higher class to a lower
-                *     class, for example split the results of a texture fetch
-                *     into individual scalar values;  We skip over these from
-                *     a 'def' perspective, and for a 'use' we walk the chain
-                *     up to the defining instruction.
-                *
-                * collect: used to collect values from lower class and assemble
-                *     them together into a higher class, for example arguments
-                *     to texture sample instructions;  We consider these to be
-                *     defined at the earliest collect source.
-                *
-                * Most of this is handled in the get_definer() helper.
-                *
-                * In either case, we trace the instruction back to the original
-                * definer and consider that as the def/use ip.
-                */
-
                if (writes_gpr(instr)) {
                        struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
                        struct ir3_register *dst = instr->regs[0];
@@ -772,27 +807,40 @@ ra_block_compute_live_ranges(struct ir3_ra_ctx *ctx, struct ir3_block *block)
                                        unsigned name = arr->base + dst->array.offset;
                                        def(name, instr);
                                }
-
                        } else if (id->defn == instr) {
-                               unsigned name = ra_name(ctx, id);
-
-                               /* since we are in SSA at this point: */
-                               debug_assert(!BITSET_TEST(bd->use, name));
-
-                               def(name, id->defn);
+                               /* in scalar pass, we aren't considering virtual register
+                                * classes, ie. if an instruction writes a vec2, then it
+                                * defines two different scalar register names.
+                                */
+                               unsigned n = ctx->scalar_pass ? dest_regs(instr) : 1;
+                               for (unsigned i = 0; i < n; i++) {
+                                       unsigned name = scalar_name(ctx, instr, i);
+
+                                       /* tex instructions actually have a wrmask, and
+                                        * don't touch masked out components.  We can't do
+                                        * anything useful about that in the first pass,
+                                        * but in the scalar pass we can realize these
+                                        * registers are available:
+                                        */
+                                       if (ctx->scalar_pass && is_tex_or_prefetch(instr) &&
+                                                       !(instr->regs[0]->wrmask & (1 << i)))
+                                               continue;
 
-                               if ((instr->opc == OPC_META_INPUT) && first_non_input)
-                                       use(name, first_non_input);
+                                       def(name, instr);
 
-                               if (is_high(id->defn)) {
-                                       ra_set_node_class(ctx->g, name,
-                                                       ctx->set->high_classes[id->cls - HIGH_OFFSET]);
-                               } else if (is_half(id->defn)) {
-                                       ra_set_node_class(ctx->g, name,
-                                                       ctx->set->half_classes[id->cls - HALF_OFFSET]);
-                               } else {
-                                       ra_set_node_class(ctx->g, name,
-                                                       ctx->set->classes[id->cls]);
+                                       if ((instr->opc == OPC_META_INPUT) && first_non_input)
+                                               use(name, first_non_input);
+
+                                       if (is_high(instr)) {
+                                               ra_set_node_class(ctx->g, name,
+                                                               ctx->set->high_classes[id->cls - HIGH_OFFSET]);
+                                       } else if (is_half(instr)) {
+                                               ra_set_node_class(ctx->g, name,
+                                                               ctx->set->half_classes[id->cls - HALF_OFFSET]);
+                                       } else {
+                                               ra_set_node_class(ctx->g, name,
+                                                               ctx->set->classes[id->cls]);
+                                       }
                                }
                        }
                }
@@ -804,7 +852,7 @@ ra_block_compute_live_ranges(struct ir3_ra_ctx *ctx, struct ir3_block *block)
                                arr->start_ip = MIN2(arr->start_ip, instr->ip);
                                arr->end_ip = MAX2(arr->end_ip, instr->ip);
 
-                               /* indirect read is treated like a read fromall array
+                               /* indirect read is treated like a read from all array
                                 * elements, since we don't know which one is actually
                                 * read:
                                 */
@@ -813,6 +861,7 @@ ra_block_compute_live_ranges(struct ir3_ra_ctx *ctx, struct ir3_block *block)
                                        for (i = 0; i < arr->length; i++) {
                                                unsigned name = arr->base + i;
                                                use(name, instr);
+                                               BITSET_SET(bd->use, name);
                                        }
                                } else {
                                        unsigned name = arr->base + reg->array.offset;
@@ -823,6 +872,37 @@ ra_block_compute_live_ranges(struct ir3_ra_ctx *ctx, struct ir3_block *block)
                                        BITSET_SET(bd->use, name);
                                        debug_assert(reg->array.offset < arr->length);
                                }
+                       } else if (ctx->scalar_pass) {
+                               struct ir3_instruction *src = reg->instr;
+                               /* skip things that aren't SSA: */
+                               unsigned n = src ? dest_regs(src) : 0;
+
+                               /* in scalar pass, we aren't considering virtual register
+                                * classes, ie. if an instruction writes a vec2, then it
+                                * defines two different scalar register names.
+                                *
+                                * We need to traverse up thru collect/split to find the
+                                * actual non-meta instruction names for each of the
+                                * components:
+                                */
+                               for (unsigned i = 0; i < n; i++) {
+                                       /* Need to filter out a couple special cases, ie.
+                                        * writes to a0.x or p0.x:
+                                        */
+                                       if (!writes_gpr(src))
+                                               continue;
+
+                                       /* split takes a src w/ wrmask potentially greater
+                                        * than 0x1, but it really only cares about a single
+                                        * component.  This shows up in splits coming out of
+                                        * a tex instruction w/ wrmask=.z, for example.
+                                        */
+                                       if (ctx->scalar_pass && (instr->opc == OPC_META_SPLIT) &&
+                                                       !(i == instr->split.off))
+                                               continue;
+
+                                       use(scalar_name(ctx, src, i), instr);
+                               }
                        } else if ((src = ssa(reg)) && writes_gpr(src)) {
                                unsigned name = ra_name(ctx, &ctx->instrd[src->ip]);
                                use(name, instr);
@@ -930,6 +1010,19 @@ ra_add_interference(struct ir3_ra_ctx *ctx)
                        debug_printf("  start_ip: %u\n", arr->start_ip);
                        debug_printf("  end_ip:   %u\n", arr->end_ip);
                }
+               debug_printf("INSTRUCTION VREG NAMES:\n");
+               foreach_block (block, &ctx->ir->block_list) {
+                       foreach_instr (instr, &block->instr_list) {
+                               if (!ctx->instrd[instr->ip].defn)
+                                       continue;
+                               debug_printf("%04u: ", scalar_name(ctx, instr, 0));
+                               ir3_print_instr(instr);
+                       }
+               }
+               debug_printf("ARRAY VREG NAMES:\n");
+               foreach_array (arr, &ctx->ir->array_list) {
+                       debug_printf("%04u: arr%u\n", arr->base, arr->id);
+               }
        }
 
        /* extend start/end ranges based on livein/liveout info from cfg: */
@@ -1060,16 +1153,34 @@ reg_assign(struct ir3_ra_ctx *ctx, struct ir3_register *reg,
 
                reg->flags &= ~IR3_REG_ARRAY;
        } else if ((id = &ctx->instrd[instr->ip]) && id->defn) {
-               unsigned name = ra_name(ctx, id);
+               unsigned first_component = 0;
+
+               /* Special case for tex instructions, which may use the wrmask
+                * to mask off the first component(s).  In the scalar pass,
+                * this means the masked off component(s) are not def'd/use'd,
+                * so we get a bogus value when we ask the register_allocate
+                * algo to get the assigned reg for the unused/untouched
+                * component.  So we need to consider the first used component:
+                */
+               if (ctx->scalar_pass && is_tex_or_prefetch(id->defn)) {
+                       unsigned n = ffs(id->defn->regs[0]->wrmask);
+                       debug_assert(n > 0);
+                       first_component = n - 1;
+               }
+
+               unsigned name = scalar_name(ctx, id->defn, first_component);
                unsigned r = ra_get_node_reg(ctx->g, name);
                unsigned num = ctx->set->ra_reg_to_gpr[r] + id->off;
 
                debug_assert(!(reg->flags & IR3_REG_RELATIV));
 
+               debug_assert(num >= first_component);
+
                if (is_high(id->defn))
                        num += FIRST_HIGH_REG;
 
-               reg->num = num;
+               reg->num = num - first_component;
+
                reg->flags &= ~IR3_REG_SSA;
 
                if (is_half(id->defn))
@@ -1077,6 +1188,16 @@ reg_assign(struct ir3_ra_ctx *ctx, struct ir3_register *reg,
        }
 }
 
+/* helper to determine which regs to assign in which pass: */
+static bool
+should_assign(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr)
+{
+       if ((instr->opc == OPC_META_SPLIT) ||
+                       (instr->opc == OPC_META_COLLECT))
+               return !ctx->scalar_pass;
+       return ctx->scalar_pass;
+}
+
 static void
 ra_block_alloc(struct ir3_ra_ctx *ctx, struct ir3_block *block)
 {
@@ -1084,20 +1205,45 @@ ra_block_alloc(struct ir3_ra_ctx *ctx, struct ir3_block *block)
                struct ir3_register *reg;
 
                if (writes_gpr(instr)) {
-                       reg_assign(ctx, instr->regs[0], instr);
-                       if (instr->regs[0]->flags & IR3_REG_HALF)
-                               fixup_half_instr_dst(instr);
+                       if (should_assign(ctx, instr)) {
+                               reg_assign(ctx, instr->regs[0], instr);
+                               if (instr->regs[0]->flags & IR3_REG_HALF)
+                                       fixup_half_instr_dst(instr);
+                       }
                }
 
                foreach_src_n(reg, n, instr) {
                        struct ir3_instruction *src = reg->instr;
+
+                       if (src && !should_assign(ctx, src) && !should_assign(ctx, instr))
+                               continue;
+
+                       if (src && should_assign(ctx, instr))
+                               reg_assign(ctx, src->regs[0], src);
+
                        /* Note: reg->instr could be null for IR3_REG_ARRAY */
                        if (src || (reg->flags & IR3_REG_ARRAY))
                                reg_assign(ctx, instr->regs[n+1], src);
+
                        if (instr->regs[n+1]->flags & IR3_REG_HALF)
                                fixup_half_instr_src(instr);
                }
        }
+
+       /* We need to pre-color outputs for the scalar pass in
+        * ra_precolor_assigned(), so we need to actually assign
+        * them in the first pass:
+        */
+       if (!ctx->scalar_pass) {
+               struct ir3_instruction *in, *out;
+
+               foreach_input (in, ctx->ir) {
+                       reg_assign(ctx, in->regs[0], in);
+               }
+               foreach_output (out, ctx->ir) {
+                       reg_assign(ctx, out->regs[0], out);
+               }
+       }
 }
 
 /* handle pre-colored registers.  This includes "arrays" (which could be of
@@ -1120,6 +1266,9 @@ ra_precolor(struct ir3_ra_ctx *ctx, struct ir3_instruction **precolor, unsigned
                        if (id->off > 0)
                                continue;
 
+                       if (ctx->scalar_pass && !should_assign(ctx, instr))
+                               continue;
+
                        /* 'base' is in scalar (class 0) but we need to map that
                         * the conflicting register of the appropriate class (ie.
                         * input could be vec2/vec3/etc)
@@ -1146,6 +1295,10 @@ ra_precolor(struct ir3_ra_ctx *ctx, struct ir3_instruction **precolor, unsigned
        }
 
        /* pre-assign array elements:
+        *
+        * TODO this is going to need some work for half-precision.. possibly
+        * this is easier on a6xx, where we can just divide array size by two?
+        * But on a5xx and earlier it will need to track two bases.
         */
        foreach_array (arr, &ctx->ir->array_list) {
                unsigned base = 0;
@@ -1223,6 +1376,58 @@ retry:
        }
 }
 
+static void
+precolor(struct ir3_ra_ctx *ctx, struct ir3_instruction *instr)
+{
+       struct ir3_ra_instr_data *id = &ctx->instrd[instr->ip];
+       unsigned n = dest_regs(instr);
+       for (unsigned i = 0; i < n; i++) {
+               /* tex instructions actually have a wrmask, and
+                * don't touch masked out components.  So we
+                * shouldn't precolor them::
+                */
+               if (is_tex_or_prefetch(instr) &&
+                               !(instr->regs[0]->wrmask & (1 << i)))
+                       continue;
+
+               unsigned name = scalar_name(ctx, instr, i);
+               unsigned regid = instr->regs[0]->num + i;
+
+               if (instr->regs[0]->flags & IR3_REG_HIGH)
+                       regid -= FIRST_HIGH_REG;
+
+               unsigned vreg = ctx->set->gpr_to_ra_reg[id->cls][regid];
+               ra_set_node_reg(ctx->g, name, vreg);
+       }
+}
+
+/* pre-color non-scalar registers based on the registers assigned in previous
+ * pass.  Do this by looking actually at the fanout instructions.
+ */
+static void
+ra_precolor_assigned(struct ir3_ra_ctx *ctx)
+{
+       debug_assert(ctx->scalar_pass);
+
+       foreach_block (block, &ctx->ir->block_list) {
+               foreach_instr (instr, &block->instr_list) {
+
+                       if ((instr->opc != OPC_META_SPLIT) &&
+                                       (instr->opc != OPC_META_COLLECT))
+                               continue;
+
+                       precolor(ctx, instr);
+
+                       struct ir3_register *src;
+                       foreach_src (src, instr) {
+                               if (!src->instr)
+                                       continue;
+                               precolor(ctx, src->instr);
+                       }
+               }
+       }
+}
+
 static int
 ra_alloc(struct ir3_ra_ctx *ctx)
 {
@@ -1236,20 +1441,54 @@ ra_alloc(struct ir3_ra_ctx *ctx)
        return 0;
 }
 
-int ir3_ra(struct ir3_shader_variant *v, struct ir3_instruction **precolor, unsigned nprecolor)
+static int
+ir3_ra_pass(struct ir3_shader_variant *v, struct ir3_instruction **precolor,
+               unsigned nprecolor, bool scalar_pass)
 {
        struct ir3_ra_ctx ctx = {
                        .v = v,
                        .ir = v->ir,
                        .set = v->ir->compiler->set,
+                       .scalar_pass = scalar_pass,
        };
        int ret;
 
        ra_init(&ctx);
        ra_add_interference(&ctx);
        ra_precolor(&ctx, precolor, nprecolor);
+       if (scalar_pass)
+               ra_precolor_assigned(&ctx);
        ret = ra_alloc(&ctx);
        ra_destroy(&ctx);
 
        return ret;
 }
+
+int
+ir3_ra(struct ir3_shader_variant *v, struct ir3_instruction **precolor,
+               unsigned nprecolor)
+{
+       int ret;
+
+       /* First pass, assign the vecN (non-scalar) registers: */
+       ret = ir3_ra_pass(v, precolor, nprecolor, false);
+       if (ret)
+               return ret;
+
+       if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
+               printf("AFTER RA (1st pass):\n");
+               ir3_print(v->ir);
+       }
+
+       /* Second pass, assign the scalar registers: */
+       ret = ir3_ra_pass(v, precolor, nprecolor, true);
+       if (ret)
+               return ret;
+
+       if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
+               printf("AFTER RA (2nd pass):\n");
+               ir3_print(v->ir);
+       }
+
+       return ret;
+}